Feature Request: GPU implementation option.

[k-a-mendoza]

The FFT implemented by numpy is pretty fast, but for some applications it is not preferable to use a CPU (even vectorized and parallel) implementation. Ex: when most GPU cores are taken up with other concurrent dask processes.

There exist a set of GPU python-CUDA bindings provided by several packages out there, so adding the option to do rfft, irfft, fft, and ifft operations may be standardized already.